codex-research

name	codex-research
description	Run a deep Codex (OpenAI) headless research pass using gpt-5.4 with full-auto mode. Use this skill whenever the user asks for "codex research", "openai research", "gpt research", "get a second opinion on the theory", or /codex-research. Sends a research question to OpenAI's Codex CLI for cross-model literature and algorithm research. Produces a written research report.

Codex Research

Run a headless research pass via the Codex CLI (codex exec) using OpenAI's strongest model. Produces a written research report and prints the results.

Usage

The user provides a research question or topic. Codex is prompted to research the topic thoroughly and write findings to a file.

TIMESTAMP=$(date +%Y%m%dT%H%M%S)
TOPIC="$1"  # brief slug for the filename
OUT_FILE=".tmp/codex-research-${TOPIC}-${TIMESTAMP}.md"
RAW_LOG="/tmp/codex-research-${TIMESTAMP}.raw.log"
mkdir -p .tmp

codex exec \
  --full-auto \
  -m gpt-5.4 \
  "${RESEARCH_PROMPT}" \
  2>&1 | tee "$RAW_LOG"

The RESEARCH_PROMPT should instruct Codex to:

1. Research the topic using its training knowledge
2. Write findings to $OUT_FILE with sections: Prior Art, Algorithms, Trade-offs, Recommendations
3. Cite specific papers, libraries, or implementations by name
4. Be explicit about what is established theory vs. speculation
5. Focus on practical applicability, not just theoretical interest

Execution

1. Construct a detailed research prompt from the user's question
2. Run the codex command in the background (run_in_background: true)
3. Wait for completion
4. Read the output file ($OUT_FILE)
5. If the output file is empty or missing, fall back to reading the raw log
6. Present the research findings to the user

Flags reference

Flag	Purpose
--full-auto	No approval prompts, workspace-write sandbox
-m gpt-5.4	OpenAI's strongest model (user's max plan)
--ephemeral	Optional: skip session persistence for faster execution

Do NOT use -o — same issue as codex-review, capture stdout instead.

Important

• Always run in the background — Codex research takes 1-3 minutes
• Always write output to .tmp/codex-research-*.md so results persist
• The --full-auto flag is critical — without it, Codex blocks waiting for approval
• If Codex fails with a model error, fall back to the default (omit -m)